Audio denoise

Recording audio in an uncontrolled environment almost always captures background noise: HVAC systems, street noise, laptop fans, keyboard clicks, air conditioning hum, electrical hiss. This noise sits in every second of the recording and is audible in quiet moments between words or during pauses in music.

Noise removal is standard post-production practice for podcasts, voiceovers, video essays, interview recordings, and any audio intended for broadcast or professional distribution. A clean audio track makes the spoken content easier to understand and removes the subliminal impression of low production quality that consistent background noise creates.

The tool uses FFmpeg's afftdn (adaptive FFT denoiser) filter, which operates in the frequency domain. The filter parameters are afftdn=nr=N:nf=-25, where nr is the noise reduction strength (higher means more aggressive reduction) and nf=-25 is the noise floor in dBFS (the threshold below which signal is treated as noise). Three strength levels are available, corresponding to different nr values.

For video files, the video stream is copied without re-encoding (-c:v copy) and only the audio stream is processed and re-encoded to AAC. For audio-only files, the output uses libmp3lame at -q:a 2 (near-lossless VBR quality). Both paths use the same afftdn filter with the selected strength.

Start with the Mild setting and increase only if the noise is not removed. Higher noise reduction settings can cause audio artifacts called musical noise: a metallic, watery quality that sounds worse than the original noise in some cases. Mild to Moderate settings are appropriate for most room noise and HVAC hum. Strong settings are appropriate for heavy electrical interference or extremely noisy recordings.

The afftdn filter is a broadband denoiser that reduces all frequencies below the noise floor. It works best on stationary noise (consistent hum, hiss, or fan noise). Variable noise (intermittent traffic, footsteps, voices in background) is much harder to remove without affecting the target audio.

Processing a video file is slower than processing audio alone because the tool must also mux the untouched video stream back into the output.

Audacity has a noise reduction effect that works by profiling a noise sample and subtracting that profile from the full recording. It is more targeted but requires identifying a clean noise section to profile. Adobe Podcast's AI enhancement uploads audio to Adobe's servers. Krisp and similar apps work in real time during calls but not on recorded files. This tool runs afftdn locally in the browser with no upload and produces a downloadable output in a single pass.